Adapting machine-learning estimators to account for the downward bias of reported greenhouse gas Scope 3 emissions

Companies are facing increased reporting demands regarding greenhouse gas (GHG) emissions as climate change evolves into a material topic for them. However, the estimation of corporate Scope 3 GHG emissions (S3E) remains challenging up to date. At the same time, we recognize the potential of machine-learning algorithms in the measurement and reduction of corporate GHG emissions. We empirically replicate the downward bias of reported S3E against true S3E. Based on such insight, it does not appear helpful to minimize the prediction risk of machine-learning estimators. Instead of point estimates, we derive that asymmetric confidence intervals based on estimators with a low bias–standard-deviation ratio (bias–SD ratio) are worthwhile to approximate true S3E from below. We develop a machine-learning algorithm with a bias–SD ratio at least 82% smaller than the bias–SD ratio of comparable methods. • Machine-learning estimators can assist companies in fulfilling forthcoming reporting obligations. • Company-reported and estimated figures tend to undermatch true Scope 3 emissions. • Mean squared error minimization is not suitable for Scope 3 estimation. • We propose estimators with a low bias–variance ratio as a potential solution.